Region of Interest Compression Method Based on Multi-layer Hyper-prior Network

Abstract

The present invention discloses a region of interest compression method based on a multi-layer hyper-prior network. By constructing and training a multi-layer hyper-prior network, the image features can be approximated to a normal distribution with a mean μ=0 through the multi-layer hyper-prior network. The spatial redundancy of the image can be fully squeezed out, overcoming the problem of incomplete redundancy removal using only a single-layer hyper-prior network in the prior art, making the present invention more widely applicable to remote sensing image compression. The attention sub-network based on the region of interest constructed by the present invention utilizes the characteristics of remote sensing images to extract the region of interest in the image as features, and can enhance the non-cloud areas of concern at different scales, thereby achieving better reconstruction performance of the region of interest, and being able to use more bit rate for compression of the region of interest. The image compression effect is improved through a precise key information enhancement mechanism.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology, and further relates to a region of interest compression method based on a multi-layer hyper-prior network in the field of remote sensing image processing technology. The present invention can be used for compressing and decompressing remote sensing images. Background Art

[0002] With the continuous reform of parallel computing of hardware devices such as GPU, the field of artificial intelligence has achieved new development, and image processing technology based on deep learning has also become an important method for multimedia data processing. Image compression technology based on region of interest (ROI) can better meet people's visual needs. When there are target individuals in the image, different compression strategies are adopted in different regions to achieve higher compression efficiency and better visual reconstruction effects. For remote sensing images, the image details are richer than ordinary images, and it is difficult to have good compression performance when the wireless channel environment is relatively limited. Therefore, in order to improve people's visual sensory experience of compressed and reconstructed images, it is necessary to perform differential compression on the region of interest and the background area.

[0003] Sung Jin Hwang et al. proposed a method for compressing the region of interest of remote sensing images based on a hyper-prior model in their paper “Variational Image Compression With AScale Hyperprior” (Published as a conference paper at ICLR 2018). The implementation steps of this method are as follows: (1) The input image is transformed through an analysis transform to obtain the potential mapping of the image. The analysis transform has four convolutional layers and three GDNs. The three convolutional layers here are the same convN×5×5 / 2↓ downsampling. (2) The potential features are input to the hyper-analysis module, which has two convN×5×5 / 2↓ and one convN×5×5 / 1↓. The activation function is Relu, and the output obtains the distribution of the image. Since it has passed through the analysis transform layer and the hyper-analysis transform layer, the output here defaults to a standard Gaussian distribution. (3) The compressed bitstream is passed through the synthesis transform module and input through four convN×5×5 / 2↑. The problem with this method is that it uses a single-layer hyper-prior and a spatial context model to promote probability estimation. This method cannot squeeze out the spatial redundancy of the image. The image features behind the single-layer hyper-prior module cannot be well approximated as a normal distribution, and the redundancy of the network is not effectively reduced.

[0004] Hangzhou Pike Pike Technology Co., Ltd. disclosed a deep learning-based encoding and decoding method in its patent application "Image encoding and decoding system and encoding and decoding method based on deep learning" (application number: 2019107059047 application date: 2019.08.01 application publication number: CN110602494 A). The steps of this method are: step 1: input image through the forward transformation network module to obtain feature coefficients; step 2: use the super prior analysis module to obtain super prior feature values; step 3: through entropy encoding and entropy decoding modules; step 4: use the reconstruction module to obtain the conditional probability model; step 5: the inverse transformation network module is used to reconstruct the image pixel value. The shortcoming of this method is that the image features are treated uniformly after the super prior analysis module and entropy encoding and decoding, and the features of different positions of the image are not considered (such as an image fully compresses the area of ​​interest, and the other parts are compressed normally). Summary of the invention

[0005] The purpose of the present invention is to address the defects of the above-mentioned technology and propose a region of interest compression method based on a multi-layer hyper-prior network to solve the problem of inability to effectively squeeze out spatial redundancy of images and effectively extract features in image compression.

[0006] The technical idea to achieve the purpose of the present invention is: the present invention constructs a multi-layer hyper-prior network, uses a multi-layer hyper-prior network, and utilizes the characteristics of image features that can be approximated to a normal distribution with a mean μ=0 through the multi-layer hyper-prior network, which can fully squeeze out the spatial redundancy of the image. In order to squeeze out the image redundancy as much as possible, reduce the spatial dependency between pixels. The input image X is analyzed and transformed to the hidden layer representation Y, and then rounded and quantized. The probability prediction network constructs an entropy model through Y calculation To estimate the likelihood Similar to the hyper-prior model. Assuming that the conditional distribution of each element in the feature representation satisfies the Gaussian distribution m, it is necessary to predict the mean μ and variance σ of the Gaussian distribution. Because the hidden layer representation has been rounded to discrete form, the likelihood of the hidden layer representation can be calculated as follows:

[0007]

[0008] Among them, Φ represents the cumulative function of the standard normal distribution, and Predicted from Y.

[0009] First, the input image is passed through the saliency detection network to obtain a binary mask. In order to obtain the binary ROI of the object area of ​​the remote sensing image, the output needs to be negated. Multiply the binary ROI and the original image to extract the non-cloud area in the image. Apply the spatial attention mechanism to the object image with the cloud area eliminated, so as to learn the importance normalized weight of the image spatial position information, multiply the weight and the unprocessed original image to be compressed, and obtain the image information after the spatial position information is enhanced. The present invention compresses the region of interest in order to focus on specific features in the remote sensing image and automatically ignore other features. By utilizing the specific characteristics of the remote sensing image, feature extraction is performed on the region of interest in the data set, and the desired features can be accurately extracted, so that more bit rate can be used for compression of the region of interest. Through the precise key information enhancement mechanism, the image compression effect is improved, and the compression of the image of the region of interest in the remote sensing image is achieved.

[0010] In order to achieve the above purpose, the technical solution adopted by the present invention includes the following steps:

[0011] Step 1: Generate training set:

[0012] Step 1.1, select at least 80 color remote sensing images with a resolution of 1000x1000 to form a sample set;

[0013] Step 1.2, input the color remote sensing images in the sample set into the trained salient region detection network in sequence, and output the binary image corresponding to the color remote sensing image;

[0014] Step 1.3, all binary images form a training set;

[0015] Step 2: Build a multi-layer hyper-prior network:

[0016] Step 2.1, build an analysis transformation subnetwork, whose structure is: first boundary filling layer, first convolution layer, first GDN layer, second boundary filling layer, second convolution layer, second GDN layer, third boundary filling layer, third convolution layer, third GDN layer, fourth boundary filling layer, fourth convolution layer;

[0017] The convolution kernel sizes of the first to fourth convolutional layers are all set to 5*5, and the convolution step sizes are all set to 2;

[0018] The first to third GDN layers are implemented by generalized normalization, whose parameters are set to β min =10 -6 , γ init =0.1, b=2 -18 ;

[0019] Step 2.2, build a synthetic transformation subnetwork, whose structure is: the first boundary filling layer, the first deconvolution layer, the first IGND layer, the second boundary filling layer, the second deconvolution layer, the second IGDN layer, the third boundary filling layer, the third deconvolution layer, and the third IGDN layer;

[0020] The convolution kernel size of the first to third deconvolution layers is set to 5*5, the step size of the deconvolution layer is set to 2, and the output boundary padding of each deconvolution layer is set to 2 layers;

[0021] The first to third IGDN layers are all implemented by inverse generalized normalization, and their parameter settings are the same as those of GDN;

[0022] Step 2.3, build two first super-prior analysis sub-networks and second super-prior analysis sub-networks with exactly the same structure, and the structure of each super-prior analysis sub-network is: first convolution layer, second convolution layer, first ReLU activation layer, third convolution layer, second ReLU activation layer, fourth convolution layer;

[0023] The size of the convolution kernel of the first convolution layer is set to 3*3, the sizes of the convolution kernels of the second to fourth convolution layers are all set to 1*1, and the step sizes of the first to fourth convolution layers are all set to 1;

[0024] The first and second ReLU activation layers are both implemented using the ReLU function to detect the features of the input samples;

[0025] Step 2.4, build two first super-prior synthesis transformation sub-networks and second super-prior synthesis transformation sub-networks with exactly the same structure, and the structure of each super-prior synthesis transformation sub-network is: first deconvolution layer, second deconvolution layer, ReLU activation layer, third deconvolution layer, ReLU activation layer, fourth deconvolution layer;

[0026] The convolution kernel sizes of the first to third deconvolution layers are all set to 1*1, the deconvolution kernel size of the fourth deconvolution layer is set to 3*3, the step size of the first deconvolution layer is set to 2, and the step sizes of the second to fourth deconvolution layers are all set to 1;

[0027] The RelU activation layer is implemented using the ReLU function to detect the features of the input samples;

[0028] Step 2.5, build an attention sub-network based on the region of interest, whose structure is: maximum pooling layer, average pooling layer, first convolutional layer, second convolutional layer, sigmoid activation layer;

[0029] The pooling kernel size of the maximum pooling layer and the average pooling layer is set to 3*3, and the moving step size of the pooling kernel is set to 1*1;

[0030] The convolution kernel size of the first and second convolutional layers is set to 3*3, and the convolution step size is set to 1;

[0031] The sigmoid activation layer is implemented using the sigmoid function, which is used to calculate the probability of the input sample being detected as each class;

[0032] Step 2.6, sequentially connect the analysis transformation sub-network, the first super-prior analysis sub-network, the second super-prior analysis sub-network, the first super-prior synthesis transformation sub-network, the second super-prior synthesis transformation sub-network, and the synthesis transformation sub-network in series to obtain a series network; the attention sub-network is respectively connected in parallel with the analysis transformation sub-network, the first super-prior analysis sub-network, and the second super-prior analysis sub-network in the series network to form a multi-layer super-prior network;

[0033] Step 3: Train a multi-layer hyper-prior network:

[0034] The binary images in the training set are input into the multi-layer super-prior network, and the back-propagation gradient descent method is used to iteratively update the parameters of each layer of the multi-layer super-prior network until the loss function of the multi-layer super-prior network converges, thereby obtaining a trained multi-layer super-prior network.

[0035] Step 4: compress and decompress the remote sensing image:

[0036] Step 4.1, input the remote sensing image to be compressed into the trained salient region detection network, and output the corresponding binary image;

[0037] Step 4.2 inputs the binary image in the data set into the trained multi-layer super-prior network, outputs the compressed code stream to be detected through the second super-prior analysis transformation sub-network, and restores the compressed code stream to the pixels of the original remote sensing image through the synthesis transformation sub-network, completing the remote sensing image compression and decompression.

[0038] Compared with the prior art, the present invention has the following advantages:

[0039] First, since the present invention constructs a multi-layer hyper-prior network, the input image features of the multi-layer hyper-prior network are approximately normally distributed, which overcomes the problem of incomplete redundancy removal using only a single-layer hyper-prior network in the prior art, making the present invention more widely used in remote sensing image compression scenarios.

[0040] Second, since the present invention constructs an attention sub-network based on the region of interest, the non-cloud area of ​​interest can be enhanced at different scales, thereby achieving better reconstruction performance of the region of interest, overcoming the problem in the prior art that the system performs unified feature extraction on the entire image when there is no region of interest, so that the present invention can maintain good performance when compressing complex remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Attached Figure 1 is a flow chart of the present invention;

[0042] Attached Figure 2 Schematic diagram of the multi-layer hyper-prior network structure of the present invention. DETAILED DESCRIPTION

[0043] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments.

[0044] Reference Figure 1 The steps for implementing the present invention are further described in detail with reference to the accompanying drawings and embodiments.

[0045] Step 1: Generate training set and test set.

[0046] Step 1.1, select 80 color remote sensing images with a resolution of 1000x1000 to form a sample set. In order to make the network generalize, the sample set is required to include 10 bands and 7 types of targets: shadow, water shadow, water, ice and snow, land, cloud and flooded area.

[0047] Step 1.2, input the color remote sensing images in the sample set into the trained salient region detection network in turn, and output the binary mask corresponding to the pixel points in each color remote sensing image, where the mask value of the cloud area is 1 and the mask value of the non-cloud area is 0. After inverting each binary mask, multiply it with the pixel point at the corresponding position in the input color remote sensing image, and replace the pixel point in the color remote sensing image with the multiplied pixel point to obtain the binary image of the color remote sensing image, and all binary images form a training set.

[0048] The trained salient region detection network is referred to as the salient region detection network proposed by Jing Zhang et al. in their paper “Deep network based on up and down blocks using wavelet transform and successful multi-scale spatial attention for cloud detection” (Remote Sensing of Environment, 261).

[0049] In the embodiment of the present invention, since the image compression task is different from the semantic segmentation task, the number of training data sets required is not as high as that of the semantic segmentation task. At the same time, in order to better subjectively compare the reconstruction and detail loss of image compression and decompression, the visible light color remote sensing image collected by Landsat is used as the sample set. The sample set is taken by the Landsat-8OLI / TIRS satellite, and the sample set is processed in step 1.2 to form the training set of the entire network.

[0050] The Landsat series of satellites is the collective name for eight satellites launched by the United States since 1972. The Landsat-8 satellite is equipped with two image acquisition payloads, namely the Operational Land Imager (OLI) and the Thermal Infrared Sensor (TIRS). The entire sample set contains at least 80 Landsat-8 OLI / TIRS data with a resolution of 1000*1000, with a total of ten bands as shown in Table 1.

[0051] Table 1 Dataset band list

[0052]

[0053] Step 2: Build a multi-layer hyper-prior network.

[0054] Step 2.1, build an analysis transformation subnetwork, whose structure is: first boundary filling layer, first convolutional layer, first GDN (Generalized Divisive Normalization) layer, second boundary filling layer, second convolutional layer, second GDN layer, third boundary filling layer, third convolutional layer, third GDN layer, fourth boundary filling layer, fourth convolutional layer.

[0055] The sizes of the convolution kernels of the first to fourth convolutional layers are all set to 5*5, the convolution step sizes are all set to 2, the first boundary padding layer is set to 2, and the second to fourth boundary padding layers are all set to 1.

[0056] The first to third GDN layers are implemented by generalized normalization, whose parameters are set to β min =10 -6 , γ init =0.1, b=2 -18 .

[0057] Step 2.2, build a synthetic transformation subnetwork, whose structure is: the first boundary filling layer, the first deconvolution layer, the first IGND (Inverse Generalized Divisive Normalization) layer, the second boundary filling layer, the second deconvolution layer, the second IGDN layer, the third boundary filling layer, the third deconvolution layer, and the third IGDN layer.

[0058] The convolution kernel sizes of the first to third deconvolution layers are all set to 5*5, the step sizes of the deconvolution layers are all set to 2, the output boundary padding of each deconvolution layer is set to 2 layers, the first to third boundary padding layers are set to 1, and the fourth boundary padding layer is set to 2.

[0059] The first to third IGDN layers are all implemented by inverse generalized normalization, and their parameter settings are the same as those of GDN.

[0060] The structures of the GDN layer and the IGDN layer are both based on existing technologies, for details, see the paper “Density Modeling of Images using a Generalized Normalization Transformation” (arXiv, 2015) published by Simoncelli EP et al.

[0061] Step 2.3, build two first super-prior analysis sub-networks and second super-prior analysis sub-networks with exactly the same structure. The structure of each super-prior analysis sub-network is: first convolutional layer, second convolutional layer, first ReLU activation layer, third convolutional layer, second ReLU activation layer, fourth convolutional layer.

[0062] The size of the convolution kernel of the first convolution layer is set to 3*3, the sizes of the convolution kernels of the second to fourth convolution layers are all set to 1*1, and the step sizes of the first to fourth convolution layers are all set to 1.

[0063] The first and second ReLU activation layers are both implemented using the ReLU function to detect the features of the input samples.

[0064] Step 2.4, build two first super-prior synthetic transformation sub-networks and second super-prior synthetic transformation sub-networks with exactly the same structure. The structure of each super-prior synthetic transformation sub-network is: first deconvolution layer, second deconvolution layer, ReLU activation layer, third deconvolution layer, ReLU activation layer, fourth deconvolution layer.

[0065] The convolution kernel sizes of the first to third deconvolution layers are all set to 1*1, the deconvolution kernel size of the fourth deconvolution layer is set to 3*3, the step size of the first deconvolution layer is set to 2, and the step sizes of the second to fourth deconvolution layers are all set to 1.

[0066] The ReLU activation layer is implemented using the ReLU function and is used to detect the features of the input samples.

[0067] Step 2.5, build an attention sub-network based on the region of interest, whose structure is: maximum pooling layer, average pooling layer, first convolutional layer, second convolutional layer, sigmoid activation layer.

[0068] The pooling kernel sizes of the maximum pooling layer and the average pooling layer are set to 3*3, and the moving step size of the pooling kernel is set to 1*1.

[0069] The convolution kernel sizes of the first and second convolutional layers are both set to 3*3, and the convolution step sizes are both set to 1.

[0070] The sigmoid activation layer is implemented using the sigmoid function, which is used to calculate the probability of the input sample being detected as each class.

[0071] Step 2.6, refer to Figure 2 To further describe the multi-layer hyper-prior network structure of the present invention.

[0072] The analysis transformation sub-network, the first super a priori analysis sub-network, the second super a priori analysis sub-network, the first super a priori synthesis transformation sub-network, the second super a priori synthesis transformation sub-network, and the synthesis transformation sub-network are sequentially connected in series to obtain a series network.

[0073] The attention sub-network is connected in parallel with the analysis transformation sub-network, the first super-prior analysis sub-network and the second super-prior analysis sub-network in the series network to form a multi-layer super-prior network.

[0074] Step 3: Train a multi-layer hyper-prior network.

[0075] The binary images in the training set are input into the multi-layer super-prior network, and the back-propagation gradient descent method is used to iteratively update the parameters of each layer of the multi-layer super-prior network until the loss function of the multi-layer super-prior network converges to obtain a trained multi-layer super-prior network.

[0076] The loss function of the multi-layer super-prior network is as follows:

[0077]

[0078] Among them, y k l represents the k-th pixel value in the l-th image, represents the predicted value of the kth pixel in the lth image, λ represents the distortion performance of controlling the bit rate, and in the embodiment of the present invention, λ is set to 0.004, M represents the total number of color remote sensing images in the training set, xi represents the i-th binary image in the training set input to the trained network, and yi is the i-th reconstructed remote sensing image output by the network. is the sum of the cross entropy of the analysis transformation subnetwork, the super prior synthesis transformation subnetwork, the super prior analysis subnetwork, and the synthesis transformation subnetwork, P X|Y is the true probability distribution of the remote sensing images in the training set, is the input distribution probability of the synthetic transformation subnetwork, is the output probability distribution of the synthetic transformation sub-network.

[0079] Step 4: compress and decompress the remote sensing image

[0080] Step 4.1, input the remote sensing image to be compressed into the trained salient region detection network, and output the corresponding binary image.

[0081] In step 4.2, the binary image is input into the trained multi-layer super-prior network, and the compressed code stream to be detected is output through the second super-prior analysis transformation sub-network. The compressed code stream is restored to the pixels of the original remote sensing image through the synthesis transformation sub-network, completing the remote sensing image compression and decompression.

[0082] The effect of the present invention is further described below in conjunction with simulation experiments:

[0083] 1. Experimental conditions:

[0084] The hardware platform of the simulation experiment of the present invention is: the processor is Intel Core i9 2080, the CPU main frequency is 3 GHz, and the memory is 16 GB.

[0085] The software platform of the simulation experiment of the present invention is: Linux operating system and python3.6.

[0086] Parameter setting: The configuration parameters of the multi-task convolutional neural network in the simulation experiment of the present invention are as follows: As described in step 1, the training set has a total of 80 images. When training the network, the optimizer uses Adam, the pixel depth (bpp) is set to 0.91, the training batch size is set to 200, the training set is trained for 5500 rounds, and the initial learning rate is set to 10 -4 .

[0087] 2. Performance evaluation and result analysis:

[0088] The network of the present invention is evaluated based on the performance indicators of the existing multi-layer hyper-prior network, including peak signal-to-noise ratio (PSNR), mean square error (MSE) and structural similarity (MS-SSIM), and the calculation formula is as follows:

[0089]

[0090] Where m represents the number of remote sensing images in the dataset, x i (i represents the i-th input image, 1≤i≤m) is the original image pixel value, y i (i represents the i-th output image, 1≤i≤m) is the pixel value of the decompressed image, μ x is x i The average value, μ y Yes i The average value of is x i The variance of Yes i The variance, σ xy is the covariance of xi and yi. c1=(k1L) 2 ,c2=(k2L) 2 It is a constant used to maintain stability. L is the dynamic range of the pixel value, which changes with the input pixel value, k1 = 0.01, k2 = 0.03, and the calculation results form Table 2, as follows:

[0091] Table 2 PSNR, MSE and MS-SSIM calculation results

[0092]

[0093] As can be seen from Table 2, the peak signal-to-noise ratio (full-image PSNR) of the full-image remote sensing image is 31.6308 and the structural similarity (full-image MS-SSIM) is 0.9855. The performance index of the region of interest consists of three parts: the peak signal-to-noise ratio of the region of interest (ROIPSNR) is 33.7148, the mean square error of the region of interest (ROI MSE) is 42.7613 and the structural similarity of the region of interest (ROI MS-SSIM) is 0.9886, which proves that the compression performance of the present invention is excellent.

[0094] The evaluation of the present invention is as follows:

[0095] The present invention designs a multi-layer super-prior image compression algorithm and a ROI spatial attention mechanism module, which enhances the network's feature extraction capability for the ROI region. It not only improves the compression performance of the overall image, but also makes better use of the bit rate for compression of the ROI region, thereby improving the image compression effect through a precise key information enhancement mechanism.

Claims

1. A region of interest compression method based on a multi-layer hyper-prior network, characterized in that: A multi-layer hyper-prior network is constructed, and the attention sub-network of the constructed region of interest enhances the non-cloud region of interest at different scales. The input image features of the multi-layer hyper-prior network are approximately normally distributed. The steps of the method include the following: Step 1: Generate training set: Step 1.1, select at least 80 color remote sensing images with a resolution of 1000x1000 to form a sample set; Step 1.2, input the color remote sensing images in the sample set into the trained salient region detection network in sequence, and output the binary image corresponding to the color remote sensing image; Step 1.3, all binary images form a training set; Step 2: Build a multi-layer hyper-prior network: Step 2.1, build an analysis transformation subnetwork, whose structure is: first boundary filling layer, first convolution layer, first GDN layer, second boundary filling layer, second convolution layer, second GDN layer, third boundary filling layer, third convolution layer, third GDN layer, fourth boundary filling layer, fourth convolution layer; The convolution kernel sizes of the first to fourth convolutional layers are all set to 5*5, and the convolution step sizes are all set to 2; The first to third GDN layers are implemented by generalized normalization, whose parameters are set to β min =10 -6 , γ init =0.1, b=2 -18 ; Step 2.2, build a synthetic transformation subnetwork, whose structure is: the first boundary filling layer, the first deconvolution layer, the first IGND layer, the second boundary filling layer, the second deconvolution layer, the second IGDN layer, the third boundary filling layer, the third deconvolution layer, and the third IGDN layer; The convolution kernel size of the first to third deconvolution layers is set to 5*5, the step size of the deconvolution layer is set to 2, and the output boundary padding of each deconvolution layer is set to 2 layers; The first to third IGDN layers are all implemented by inverse generalized normalization, and their parameter settings are the same as those of GDN; Step 2.3, build two first super-prior analysis sub-networks and second super-prior analysis sub-networks with exactly the same structure, and the structure of each super-prior analysis sub-network is: first convolution layer, second convolution layer, first ReLU activation layer, third convolution layer, second ReLU activation layer, fourth convolution layer; The size of the convolution kernel of the first convolution layer is set to 3*3, the sizes of the convolution kernels of the second to fourth convolution layers are all set to 1*1, and the step sizes of the first to fourth convolution layers are all set to 1; The first and second ReLU activation layers are both implemented using the ReLU function to detect the features of the input samples; Step 2.4, build two first super-prior synthesis transformation sub-networks and second super-prior synthesis transformation sub-networks with exactly the same structure, and the structure of each super-prior synthesis transformation sub-network is: first deconvolution layer, second deconvolution layer, ReLU activation layer, third deconvolution layer, ReLU activation layer, fourth deconvolution layer; The convolution kernel sizes of the first to third deconvolution layers are all set to 1*1, the deconvolution kernel size of the fourth deconvolution layer is set to 3*3, the step size of the first deconvolution layer is set to 2, and the step sizes of the second to fourth deconvolution layers are all set to 1; The RelU activation layer is implemented using the ReLU function to detect the features of the input samples; Step 2.5, build an attention sub-network based on the region of interest, whose structure is: maximum pooling layer, average pooling layer, first convolutional layer, second convolutional layer, sigmoid activation layer; The pooling kernel size of the maximum pooling layer and the average pooling layer is set to 3*3, and the moving step size of the pooling kernel is set to 1*1; The convolution kernel size of the first and second convolutional layers is set to 3*3, and the convolution step size is set to 1; The sigmoid activation layer is implemented using the sigmoid function, which is used to calculate the probability of the input sample being detected as each class; Step 2.6, sequentially connect the analysis transformation sub-network, the first super-prior analysis sub-network, the second super-prior analysis sub-network, the first super-prior synthesis transformation sub-network, the second super-prior synthesis transformation sub-network, and the synthesis transformation sub-network in series to obtain a series network; the attention sub-network is respectively connected in parallel with the analysis transformation sub-network, the first super-prior analysis sub-network, and the second super-prior analysis sub-network in the series network to form a multi-layer super-prior network; Step 3: Train a multi-layer hyper-prior network: The binary images in the training set are input into the multi-layer super-prior network, and the back-propagation gradient descent method is used to iteratively update the parameters of each layer of the multi-layer super-prior network until the loss function of the multi-layer super-prior network converges, thereby obtaining a trained multi-layer super-prior network. Step 4: compress and decompress the remote sensing image: Step 4.1, input the remote sensing image to be compressed into the trained salient region detection network, and output the corresponding binary image; In step 4.2, the binary image in the data set is input into the trained multi-layer super-prior network, and the compressed code stream to be detected is output through the second super-prior analysis transformation sub-network. The compressed code stream is restored to the pixels of the original remote sensing image through the synthesis transformation sub-network, completing the remote sensing image compression and decompression.

2. The method for compressing a region of interest based on a multi-layer hyper-prior network according to claim 1, characterized in that: The sample set described in step 1.1 includes at least 10 bands and seven types of targets: shadow, water shadow, water, ice and snow, land, cloud, and flooded area.

3. The region of interest compression method based on a multi-layer hyper-prior network according to claim 1, characterized in that: The binary image described in step 1.2 is obtained by the following steps: input the color remote sensing images in the sample set into the trained salient region detection network in turn, output the binary mask corresponding to the pixel points in each color remote sensing image, set the mask value of the cloud area to 1, and set the mask value of the non-cloud area to 0, invert each binary mask, and then multiply it with the pixel points at the corresponding position in the input color remote sensing image, replace the pixel points in the color remote sensing image with the multiplied pixel points, and obtain the binary image of the color remote sensing image.

4. The method for compressing a region of interest based on a multi-layer hyper-prior network according to claim 1, characterized in that: The loss function of the multi-layer hyper-prior network described in step 3 is as follows: Among them, loss represents the loss function of the multi-layer hyper-prior network, λ represents the distortion performance of controlling the bit rate, M represents the total number of color remote sensing images in the training set, and x i represents the i-th remote sensing image input to the trained multi-layer hyper-prior network, y i represents the i-th reconstructed remote sensing image output by the trained multi-layer hyper-prior network, represents the sum of the cross entropy of the analysis transformation sub-network, the super-prior synthesis transformation sub-network, the super-prior analysis sub-network, and the synthesis transformation sub-network, P X|Y represents the true probability distribution of the remote sensing images in the training set, represents the input distribution probability of the synthetic transformation sub-network, Represents the output probability distribution of the synthesis transformation sub-network.

Images